Decoupling Telephony from the AI Monolith: The Siphon Architecture
Siphon it!
If you are an engineer tasked with building a real-time voice agent, you usually face a grim choice: stitch together raw SIP trunks, WebRTC bridges, and VAD algorithms yourself (taking months), or surrender to a black-box CPaaS and pay double the raw compute costs.
At BLACKDWARF, we spent months dealing with unpredictable network jitter, Head-of-Line blocking over TCP WebSockets, and messy state management.
We realized the ecosystem needed a production-ready middleware layer.
Enter Siphon.
The Architectural Problem: State vs. Scale
Unlike stateless REST APIs, AI telephony is highly stateful. A continuous RTP media stream must be maintained between the user's phone (via SIP), the Speech-to-Text engine, the LLM, and the Text-to-Speech synthesizer.
To achieve sub-500ms latency, we abandoned standard HTTP/WebSockets entirely for the media pipeline.
Siphon is architected directly on top of LiveKit's WebRTC infrastructure.
How Siphon Routes Media
Inbound SIP: A call hits a Twilio/Telnyx number, which forwards the SIP INVITE to Siphon's dispatch node.
WebRTC Bridge: Siphon converts the raw SIP media stream into a low-latency WebRTC track.
Local VAD: We run Voice Activity Detection locally to determine exactly when the user starts and stops speaking, bypassing the latency of cloud-based VADs.
Provider Agnostic Execution: Siphon sends the text chunk to your configured LLM (OpenAI/Anthropic).
Zero-Config Horizontal Scaling
Because maintaining state across nodes is a DevOps nightmare, Siphon includes an autonomous load balancer.
You do not need Kubernetes HPA to scale Siphon.
When you run your agent worker script, the Siphon node automatically broadcasts its presence.
The master dispatcher detects new workers and organically load-balances SIP traffic across the fleet.
You can scale from 1 concurrent call on your laptop to 1,000 concurrent calls in production without changing a single line of architecture.
By decoupling the telephony infrastructure from the AI provider, Siphon gives engineers total control over their stack.
Review our Python source code and architecture diagrams on our GitHub repo to see exactly how we manage the event loops.
